Efficient adaptive learning rate for convolutional neural network based on quadratic interpolation egret swarm...

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) are widely used for multi-domain image classification.
• Existing CNN methods often exhibit suboptimal performance and convergence issues across diverse datasets.
• Adaptive learning rate optimization is crucial for enhancing prediction accuracy in deep learning models.

Purpose of the Study:

• To propose a novel algorithm for multi-domain image classification using CNNs.
• To introduce an adaptive learning rate rule to improve CNN performance and convergence.
• To enhance prediction accuracy by optimizing learning rate hyperparameters.

Main Methods:

• Utilized CNNs for extracting rich feature representations from images.
• Introduced the Egret Swarm Optimization Algorithm (ESOA) to adaptively update the learning rate, aiding in escaping local extrema.
• Incorporated quadratic interpolation to approximate the objective function, thereby boosting prediction accuracy.

Main Results:

• Achieved a highest accuracy rate of 97.15% on a test set for multi-domain image classification.
• Demonstrated superior performance of the proposed algorithm over Particle Swarm Optimization (PSO), Genetic Algorithm (GA), Whale Optimization Algorithm (WOA), Catch Fish Optimization Algorithm (CFOA), and GOOSE Algorithm (GO) on benchmark functions.
• The algorithm ranked first in performance metrics across CEC2017 and CEC2022 benchmark sets, particularly excelling on unimodal functions.

Conclusions:

• The proposed adaptive learning rate CNN algorithm significantly improves multi-domain image classification performance.
• ESOA effectively optimizes learning rates, leading to enhanced accuracy and faster convergence.
• The method shows robustness and superiority compared to existing optimization algorithms on various benchmark tasks.